// Gemini API helpers — single key, used for audio transcription, vision, and text.
//
// API key is stored in localStorage under 'pepagent.geminiKey'.
// Set/clear it with the gear button in the header (see GeminiKeyButton).
//
// All three helpers below hit the v1beta generativelanguage endpoint.
const GEMINI_MODEL = 'gemini-2.5-flash';
const GEMINI_KEY_STORAGE = 'pepagent.geminiKey';
function getGeminiKey() {
try { return localStorage.getItem(GEMINI_KEY_STORAGE) || ''; } catch (e) { return ''; }
}
function setGeminiKey(k) {
try { localStorage.setItem(GEMINI_KEY_STORAGE, k || ''); } catch (e) {}
}
// Read a Blob/File as base64 (no data: prefix)
function blobToBase64(blob) {
return new Promise((resolve, reject) => {
const r = new FileReader();
r.onload = () => {
const s = String(r.result || '');
const i = s.indexOf(',');
resolve(i >= 0 ? s.slice(i + 1) : s);
};
r.onerror = () => reject(r.error || new Error('read failed'));
r.readAsDataURL(blob);
});
}
// Two paths:
// 1. User pasted their own key in the modal → call Google directly from the browser.
// 2. No user key → call /api/gemini (server proxy with GEMINI_API_KEY env var).
//
// The proxy is the default on the deployed site. The modal is a power-user override.
async function geminiCall(parts, opts = {}) {
const { temperature = 0.7, maxOutputTokens = 1024, model, disableThinking = false } = opts;
const userKey = getGeminiKey();
const useModel = model || GEMINI_MODEL;
const generationConfig = { temperature, maxOutputTokens };
if (disableThinking) generationConfig.thinkingConfig = { thinkingBudget: 0 };
if (userKey) {
const url = `https://generativelanguage.googleapis.com/v1beta/models/${useModel}:generateContent?key=${encodeURIComponent(userKey)}`;
const res = await fetch(url, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
contents: [{ role: 'user', parts }],
generationConfig,
}),
});
if (!res.ok) {
const errText = await res.text().catch(() => '');
throw new Error(`Gemini ${res.status}: ${errText.slice(0, 200)}`);
}
const j = await res.json();
const out = j?.candidates?.[0]?.content?.parts?.map((p) => p.text || '').join('').trim();
if (!out) throw new Error('Gemini returned empty response');
return out;
}
// Proxy path
const res = await fetch('/api/gemini', {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({ parts, temperature, maxOutputTokens, model: useModel, disableThinking }),
});
if (!res.ok) {
let detail = '';
try { const j = await res.json(); detail = j.error || j.detail || ''; } catch (e) {}
throw new Error(`Proxy ${res.status}: ${detail || 'request failed'}`);
}
const j = await res.json();
if (!j.text) throw new Error('Proxy returned empty response');
return j.text;
}
// Send audio (Blob from MediaRecorder) → text transcript.
//
// Notes:
// - Strip codec params from MIME (Gemini wants "audio/webm", not "audio/webm;codecs=opus").
// - Pad short audio with a low temp so the model doesn't refuse on greetings.
// - Don't use a sentinel like [unintelligible] — the model leaks it on edge cases.
async function geminiTranscribe(audioBlob) {
const b64 = await blobToBase64(audioBlob);
const rawMime = audioBlob.type || 'audio/webm';
const mime = rawMime.split(';')[0].trim() || 'audio/webm';
return geminiCall([
{ text: 'Transcribe the spoken words in this audio clip. Output ONLY the transcript text — nothing else, no quotes, no labels, no commentary, no explanations. If you hear no clear words at all, output a single space character.' },
{ inlineData: { mimeType: mime, data: b64 } },
], {
temperature: 0.2,
maxOutputTokens: 1024,
disableThinking: true, // critical: 2.5-flash burns all tokens on thinking otherwise
});
}
// Send an image + system context → analysis.
// The agent figures out whether it's a vial/box label, a body shot, or something else,
// and responds accordingly.
async function geminiAnalyzePeptidePhoto(imageBlob, { goal, profile, protocol } = {}) {
const b64 = await blobToBase64(imageBlob);
const mime = imageBlob.type || 'image/jpeg';
const ctx = [];
if (protocol) ctx.push(`Active protocol: ${protocol.label} (${protocol.stack.join(' + ')}, ${protocol.duration}).`);
if (profile) ctx.push(`User: ${profile.weight}kg, ${profile.age}y, ${profile.experience}.`);
const ctxLine = ctx.length ? '\n\nContext: ' + ctx.join(' ') : '';
const prompt = `You are Pepagent, a peptide protocol advisor. Look at this photo and figure out what it is, then respond.
Possible subjects:
1. PEPTIDE VIAL / BOX / LABEL / RECONSTITUTION KIT — read the peptide name, dose, lot/batch, expiration, supplier if visible. Give: typical dosing for this peptide, mechanism in one sentence, half-life, common stack synergies, and any red flags you notice on the label (sketchy supplier wording, missing lot/batch, damaged seal, suspicious appearance of the powder/solution).
2. INJECTION SITE / BODY SHOT — comment on site rotation if visible, flag any concerning marks (significant redness, swelling, unusual bruising, lumps, infection signs), and remind about rotating sites every injection.
3. PROGRESS PHOTO (body composition / skin / hair) — give honest visual feedback against the user's stated goal, what to track over the cycle, and what a realistic timeline for visible change looks like.
4. SOMETHING ELSE — say briefly what you see and ask the user what they want to know.
Be conversational, 2-4 short paragraphs, no markdown, no bullet points, no headers. End with a one-line reminder that this is educational and to confirm with a clinician.${ctxLine}`;
return geminiCall([
{ text: prompt },
{ inlineData: { mimeType: mime, data: b64 } },
], { temperature: 0.6, maxOutputTokens: 1024 });
}
// Plain text generation — used as a fallback / alternative to window.claude.complete.
async function geminiText(prompt) {
return geminiCall([{ text: prompt }], { temperature: 0.7, maxOutputTokens: 1024 });
}
// Header gear button + modal to paste the key.
function GeminiKeyButton({ accent }) {
const [open, setOpen] = React.useState(false);
const [key, setKey] = React.useState(() => getGeminiKey());
const hasKey = !!getGeminiKey();
const save = () => { setGeminiKey(key.trim()); setOpen(false); };
const clear = () => { setGeminiKey(''); setKey(''); setOpen(false); };
return (
<>
{open && (
{ if (e.target === e.currentTarget) setOpen(false); }}>
// API CONFIG
Gemini API key
Used for voice transcription and photo analysis. Stored in your browser only — never sent anywhere except Google.
Get one at aistudio.google.com/app/apikey